Hunyuan Multi-Lora Loader

Every HunyuanVideo LoRA guide eventually hands you the same ugly graph: a LoraLoader node per LoRA, chained in series, each one re-wiring the model into the next. Three or four LoRAs in and the wires are a mess - and more importantly, the output starts going soft. Hunyuan Multi-Lora Loader is one small node that replaces that whole chain. It takes your model, applies up to four LoRAs in a single pass, and gives you back a model that's ready for the sampler.

This is a genuinely tiny pack from a single author, and it does one thing. Don't expect a community behind it; the point is that the one thing it does is the right one.

Why Hunyuan needed this

The README's framing is honest: Kijai's excellent ComfyUI-HunyuanVideoWrapper (the standard way to run HunyuanVideo in ComfyUI) just chains LoRAs in a way that gets fiddly once you stack several. The pack description goes further - it exists to prevent "blurriness and other artifacts" when loading multiple LoRAs in Hunyuan Video. Video LoRAs don't stack as cleanly as image LoRAs; anyone who's tried 3-4 stacked LoRAs on a video model has watched quality collapse or bleed. This node's answer is blocks_type.

HunyuanVideo's transformer isn't one uniform stack. It has two kinds of blocks: single_blocks and double_blocks (the doubles carry the text cross-attention, which is where most of what a content LoRA wants to change lives). When you set a slot to double_blocks, the node loads the LoRA weights, strips the diffusion_model. / transformer. prefixes off the key names, and keeps only the keys that belong to that block type before applying them. That's the mechanism, and it's why the author's usage tip is blunt: "Please select double_blocks."

What you actually set

Four identical slots, lora_01 through lora_04. Each one has three fields, and the middle two are what you'll touch:

• lora_01..04 - dropdown of everything in ComfyUI/models/loras. Set to "None" to disable a slot.
• strength_01..04 - default 1.0, but the range runs -10 to 10. Going negative is how you suppress a style instead of adding it.
• blocks_type_01..04 - all, single_blocks, or double_blocks. Start with double_blocks and only widen if the LoRA is being ignored.

The only other input is model (a MODEL), which comes straight from your HunyuanVideo loader. The single output is also model - wire it into your KSampler or whatever comes next. Nothing else. If you've used rgthree's Power Lora Loader for image models, this is the same idea, narrowed to HunyuanVideo and stripped of the extras.

Installing it

Two ways, both trivial:

# Option 1: ComfyUI Manager
# Manager → Install Custom Nodes → search "Hunyuan-Multi-Lora-Loader" → Install → Restart

# Option 2: manual
cd ComfyUI/custom_nodes
git clone https://github.com/asdrabael/Hunyuan-Multi-Lora-Loader
# then restart ComfyUI

There are no dependencies to install, no model files to download, no Python packages. The whole node is a single file that uses ComfyUI's built-in load_lora_for_models and load_torch_file. What you do need is the actual HunyuanVideo model on disk (Kijai's wrapper or native ComfyUI loader) and your LoRAs in the loras folder - this node assumes those exist, it doesn't fetch them.

Where people get burned

A few things worth knowing, mostly straight from the source:

• A slot does nothing if strength is 0 or the name is "None" - the node skips it entirely. Handy for building one workflow and testing LoRAs by flipping them off.
• The author's testing default is double_blocks. If a LoRA looks like it's not applied at all, that's the first thing to check - with double_blocks you're only applying part of it, which is the point, but it does mean a LoRA whose weights live mostly in single blocks will whisper.
• Don't crank all four to strength 1.0 and expect magic. Hunyuan LoRAs over-stack fast, and this node makes it easy to stack too much. Start low, add slots one at a time.

One honest caveat: this is a one-commit, one-author pack from early 2025. It works on what it was written against, it uses only stable ComfyUI APIs, and there's an IS_CHANGED handler so ComfyUI actually re-runs when you change a slot (some loaders cache and don't). But if ComfyUI's internals drift, don't expect a fast upstream fix. For a focused "clean up my Hunyuan LoRA graph" problem, it's still the node I'd reach for.